Sigma-Delta Learner for Source Separation and localization in Miniature Microphone Arrays

Acoustic sensing is one of several emerging areas where micro/nano-scale integration promises significant breakthroughs. Micro-ElectroMechanical Systems (MEMS) microphones can integrate analog-to-digital converters on the same chip to produce digital outputs:


With this technology, an ever increasing number of microphones can now be integrated within a single package in order to take advantage of microphone arrays' properties. These miniature microphone arrays have many applications including intelligent hearing aids and acoustic surveillance. It is envisioned that next generation of intelligent hearing devices will integrate hundreds of micro/nanoscale microphones, separate speech from noise by localizing different acoustic sources. It is also envisioned that surveillance robots and sensor networks will utilize the microphone array to remotely localize, track and identify speakers of interest. Different types of these microphone arrays have been implemented in AIMLab:


However, separation and localization of acoustic sources using miniature microphone arrays poses a significant challenge due to fundamental limitations imposed by the physics of sound propagation. In this project, we propose a novel framework that overcomes these limitations by integrating learning algorithms directly with the signal measurement (analog-to-digital) process and enables very high fidelity separation and localization of mixtures.

Analog to LPC Converter

For many speech/speaker recognition systems, the feature extraction unit is the most computationally intensive and power consuming component. In this project, we investigate a design of an analog-to-feature converter that directly produces a pulse-encoded representation of the linear predictive coded (LPC) features corresponding to an input analog signal. At the core of the proposed design is a sigma-delta modulation.

Analog Spectral Features

In this project we investigate use of analog spectral features for speeah/speaker recognition task. The advantage of using spectral features is that the signal flow is feed-forward and inherently stable. The analog spectral model can lead to efficient implementation in analog VLSI while offering good recognition performance. The proposed approach consists of four stages: a Mel-based band-pass filter bank, a rectification stage, a low-pass stage, and a cubic compression stage.

Kernel Filtering for Robust Speech Features

In this project, we investigate a non-linear filtering approach for extracting noise-robust speech features that can be used in a speech/speaker recognition task. At the core of the proposed approach is a time-series regression using Reproducing Kernel Hilbert Space (RKHS) based methods that extracts discriminatory non-linear signatures while filtering out the noninformative noise components.